MASS Separates World Logic From View Rendering to Run Multiplayer World Models With 1,024 Concurrent Players for 10,000 Recurrent Steps

arXiv / HuggingFace Daily Papers (7 upvotes)·low signalXBlueskyLinkedInCopy link

MASS (arXiv 2608.06257, submitted August 6) argues current video world models fail in multiplayer settings because they entangle world state with view-dependent visual latents, causing redundant compute, cross-view inconsistency and poor scaling. It splits the system into a Logic Engine that maintains authoritative global state from joint actions and a Rendering Engine that produces independent per-camera views, reporting better state accuracy and lower cross-view inconsistency than multi-view baselines on a matched multiplayer Snake benchmark. The headline scalability claim is advancing predicted worlds with 1,024 concurrent players over 10,000 recurrent steps.